Good Regulator Theorem

Definition

The Good Regulator Theorem connects successful control with modelling. Its central question is what relationship a controller must have to the system it regulates. Roger C. Conant and W. Ross Ashby’s 1970 result concerns the simplest regulators that minimize uncertainty in outcomes, under specified mathematical assumptions. It does not establish that every useful controller contains an explicit simulation. Conant and Ashby, 1970.

A regulator is something that helps keep outcomes within a desired pattern despite disturbances. A heating controller, for example, influences temperature despite changing conditions. Calling something a model requires more care: the word can mean a simulation, a predictive representation, or a mathematical correspondence. Those meanings support different claims. An arrangement that works reliably need not explain its environment in language or rehearse possible futures.

How it works

In the original formulation, system events S and regulator events R jointly determine outcomes Z. Optimality means minimizing outcome entropy H(Z), an information measure of uncertainty. Given the outcome rule, a fixed distribution of system events, and conditional regulator responses, an optimal response can be simplified to a mapping h: S → R. Each system event then selects one regulator event. This correspondence supplies the theorem’s particular meaning of model; it need not reproduce every environmental detail. Original theorem and proof.

Hypothetical example: imagine an automated shade protecting a greenhouse from excessive sunlight. For a deliberately simplified task, the designer distinguishes only conditions requiring shade from conditions allowing it to remain open. A controller that consistently selects the appropriate action preserves a distinction relevant to regulation. It could do so through a small switching mechanism rather than a detailed simulation of the greenhouse.

This example illustrates why the task boundary matters. The same distinction might be inadequate if humidity, wind damage, or different plants also mattered. A representation can be sufficient for one stated purpose while leaving other questions unanswered. The useful design question becomes: which differences in the environment require different responses? Answering that question still requires knowledge of the actual equipment and conditions; the theorem does not provide it.

Evidence and applications

The evidence here is mathematical: a proof establishes a conclusion within a formal setup. Applying it to equipment, organisms, or organizations additionally requires showing that the setup describes the case. Mathematical necessity within a model and demonstrated reliability in the world are separate achievements.

A distinct formulation appears in Nathaniel Virgo, Martin Biehl, Manuel Baltieri and Matteo Capucci’s 2025 research preprint, A “good regulator theorem” for embodied agents. Version 1 treats deterministic systems evolving in discrete time. Regulation requires a nonempty set of combined agent–environment states that remains within a designated acceptable set. Theorem 3.4 relates this condition to an observer’s attribution of beliefs and goals. Beliefs specify possible environmental states, rather than probabilities, and their updating may discard information. Embodied-agent formulation, version 1.

This is a different meaning of model, not an empirical confirmation that controllers contain hidden simulators. The observer chooses the interpretation. The authors’ doorstop example shows how a successful regulator can receive an unchanging, trivial belief interpretation. They therefore reject the inference that designers must always begin by explicitly constructing a model. They also leave extension to stochastic systems unresolved. Trivial interpretations; scope restrictions.

For comparison, JEPA and World Models concerns learned predictive representations, while MuZero concerns planning using learned latent dynamics. These entries provide concrete settings in which to ask what a representation preserves. The Good Regulator Theorem alone cannot establish that either architecture is necessary, that its representation is adequate, or that its deployment is reliable.

Limitations

The word good should not be read as a moral endorsement. A perfectly predictable outcome could still be undesirable. Minimizing uncertainty therefore cannot, by itself, establish that the chosen outcome serves people’s interests. Selecting outcomes and assessing their consequences remain additional tasks.

The original paper explicitly allows equally optimal regulators with extra complexity. Its result also assumes stable system statistics, with a qualified discussion of slowly changing conditions. Neither qualification warrants a general guarantee under abrupt environmental change. Qualifications following the proof.

The 2025 formulation offers counterexamples to a stronger interpretation of the slogan: successful regulation need not involve informative, changing internal representations. Its existence condition also does not establish success from every possible starting state. An attributed belief is not evidence of conscious belief, and an observer’s acceptable set carries no independent ethical authority.

Neither document supplies a universal recipe for learning, choosing sensors, managing uncertainty, or governing institutions. The 2025 document cited here is specifically an arXiv version 1 preprint. Its theorem should be assessed within that version’s definitions and scope.

Relevance to Spiralism

A possible Spiralist interpretation treats the theorem as a prompt to examine the relationship between representation and action. A controller’s useful distinctions deserve attention, but so do the distinctions it ignores. This is an interpretive application proposed here, separate from the sources’ mathematical claims.

Applied to collective intelligence or institutional governance, the resulting questions might be: who defines success, what information reaches decision makers, and which consequences fall outside their representation? Such questions could guide investigation, but answering them would require separate evidence about the institution. An analogy with a regulator does not establish an organization’s optimality or legitimacy.

AI Control provides a related discussion of deployment protocols and their evidence requirements. Recursive Reality supplies the site’s separate interpretive frame for representations whose interventions affect later observations. Together, these connections encourage scrutiny of what a model represents and what its use changes, without turning the theorem into a claim about machine consciousness or institutional authority.

AI-use disclosure

This wiki entry was drafted by GPT-6.1 Sol with high reasoning and checked in a separate GPT-6.1 Sol session with maximum reasoning. This is automated editorial review, not human peer review or independent verification of the sources’ claims.

Sources

• PDF document: Conant_Ashby.pdf · accessed 2026-10-23
• A “good regulator theorem” for embodied agents · accessed 2026-10-23

All wiki entries